Fast and Accurate Machine Learning Strategy for Calculating Partial Atomic Charges in Metal–Organic Frameworks

Computational high-throughput screening
using molecular simulations
is a powerful tool for identifying top-performing metal–organic
frameworks (MOFs) for gas storage and separation applications. Accurate
partial atomic charges are often required to model the electrostatic
interactions between the MOF and the adsorbate, especially when the
adsorption involves molecules with dipole or quadrupole moments such
as water and CO2. Although ab initio methods can be used
to calculate accurate partial atomic charges, these methods are impractical
for screening large material databases because of the high computational
cost. We developed a random forest machine learning model to predict
the partial atomic charges in MOFs using a small yet meaningful set
of features that represent both the elemental properties and the local
environment of each atom. The model was trained and tested on a collection
of about 320 000 density-derived electrostatic and chemical
(DDEC) atomic charges calculated on a subset of the Computation-Ready
Experimental Metal–Organic Framework (CoRE MOF-2019) database
and separately on charge model 5 (CM5) charges. The model predicts
accurate atomic charges for MOFs at a fraction of the computational
cost of periodic density functional theory (DFT) and is found to be
transferable to other porous molecular crystals and zeolites. A strong
correlation is observed between the partial atomic charge and the
average electronegativity difference between the central atom and
its bonded neighbors.